The honest answer to "how do I get past AI detection" is two ideas held together. When AI was used (research, outline, grammar, brainstorming), disclose it in a one-line note. When you did the writing, prove it with an authentic-voice draft and a dated scan history. There are no tricks on this page and no promises of permanent invisibility, because both expire the moment a detector ships an update. What you get instead is a five-step workflow that holds up under review, three honest paths depending on how AI actually fit into your process, and the disclosure language that closes most academic and editorial conversations before they start.
The ethical path is faster, more durable, and easier to defend than any score-reduction workflow. Two ideas anchor everything that follows.
If you did the thinking, the prose should sound like you. That means specific anchors from your own reading, a sentence rhythm that reflects how you actually write, and a refusal to substitute a paraphraser's voice for yours. A draft that genuinely carries your voice will usually land in the 5 to 25 percent AI band on TextSight as a side effect of authorship, not as a goal you chased. When the underlying signal is honest, no detector update reverses the result.
If AI helped on the outline, the research summary, the grammar pass, or a brainstorming session, say so in a one-line note at submission. Most institutions and editors in 2026 explicitly permit disclosed AI assistance for those uses. The disclosure costs nothing, removes ambiguity, and gives the grader or editor the information they need to evaluate your work fairly. Hidden assistance is what gets people in trouble, not disclosed assistance.
The strongest evidence that you wrote the work is not a detector score. It is the version history showing the document growing sentence by sentence, the dated TextSight scans across your revisions, your research notes, and a short timeline of when you worked on the piece. Build the trail while you write, not after a flag appears. Five minutes of habit replaces hours of anxiety later.
The workflow is almost identical across paths; the disclosure language is what changes. Be honest with yourself about which one you are on before you write a single line of revision.
You wrote the outline, the prose, and the citations yourself. AI was never in the loop. The workflow is simplest here: write the draft, scan it on TextSight to verify, and revise any sentences that flag (usually because the topic is over-saturated or your sentence rhythm is uniform, not because the work is AI). No disclosure is needed because there is nothing to disclose, and the dated scan history is your insurance against a future false-positive flag. This is the path most ESL writers, formal academic writers, and lawyers find themselves on; the work is honest and the detector is the problem.
You used ChatGPT to summarise three background sources, or asked it to brainstorm angles before you settled on one, or ran the finished draft through Grammarly for a final grammar pass. The prose is yours and the argument is yours, but AI touched the process. Write the draft yourself, scan with TextSight, revise as needed, then add a one-line disclosure at submission: "I used AI for research summaries and grammar; prose, argument, and citations are my own." That sentence removes 95 percent of grey-area conversations with graders and editors.
You let ChatGPT draft a section or two, then revised the paragraphs heavily until they carried your reasoning and your phrasing. This is the most disclosure-sensitive path and the one most academic policies treat differently. The honest move is to rewrite enough that the prose genuinely reflects your thinking (not just a synonym pass), keep version history showing the revision, and disclose specifically: "AI was used to draft two paragraphs in the literature review which I then rewrote; the rest of the prose and all citations are my own." If your venue forbids any AI-generated prose, this path is not available; switch to path one and write the section yourself.
Roughly 25 to 30 minutes on an 800-word piece once you know the patterns. The steps are identical across the three paths; only the disclosure note at step five changes.
Write a one-line note for your own records before you start revising. Was AI used for the outline, the research, the grammar pass, the summary, or not at all. Was it ChatGPT, Claude, Gemini, or a grammar tool. The honest path forks here, and naming the assistance is the prerequisite for the rest of the workflow. Most writers skip this step because it feels obvious, then realise three hours later they cannot remember exactly which paragraph came from where. Two minutes of writing-it-down replaces hours of reconstruction.
Even if the outline came from AI, type the sentences yourself. Specific anchors, your reading, your phrasing, the way you would explain the idea to a friend. The unevenness of human writing in real time is what no detector can fake because it is not a pattern; it is a sequence of choices. Three concrete moves carry a draft into authentic-voice territory: vary sentence length deliberately (one short sentence under eight words and one long sentence over 28 in each paragraph), use a contraction where conversation would call for one, and drop in a date or a named source the AI would not have invented.
Paste the draft into TextSight at app.textsight.ai. Free for three scans a day (up to 1,000 words each with no signup, or 1,500 with a free account), with a published false-positive rate; the sentence-level highlight map starts on Starter. Read the map, not just the aggregate percentage. Green sentences are fine, yellow are borderline, red are the ones the detector mistook for AI. Most honest drafts have a handful of red sentences clustered in introductions and conclusions where formal scaffolding lives, not throughout the whole piece.
Edit only the red sentences. Leave the rest alone. Five anchor types do most of the work: a date (1789 reads more human than "the late eighteenth century"), a named author you actually read, a five-to-ten-word direct quote, a counter-example from your notes, or a personal observation tied to the argument. Cut tripled adjectives ("a robust, comprehensive, multifaceted approach" becomes "an approach that handles three cases"). Break uniform rhythm by adding one short sentence and one long one per paragraph. Refuse to run honest prose through a paraphraser; that substitutes a machine voice for yours and usually drops the score.
If you are on path two or path three, add a one-line disclosure note when you submit. Keep it specific rather than generic; the goal is to let the grader or editor evaluate the role AI played without guessing. Pair the note with your dated TextSight scan history, which acts as an audit trail showing how the draft evolved. On Pro the scan history keeps every scan automatically, which is the practical reason most students upgrade. Most institutions in 2026 published guidance explicitly stating that disclosed AI assistance for outlining, research, or grammar is acceptable; the disclosure is the move that converts a grey-area situation into a clearly-permitted one.
Free includes 3 actions a day (scans or humanizations, any mix) at up to 260 words per humanization. Paid tiers raise quotas and add the Chrome extension, file upload, and REST API. Yearly billing saves 63%.
Yearly billing saves 63%. View full pricing
The humanizer is polish for the last few flagged sentences after step four, not a one-click pass over an honest draft. Pick the mode that matches the work, and run it only on the residuals.
Basic is the free default and the lighter pass, staying close to the original phrasing. It is the right choice for academic writing where exact meaning matters, for legal and technical writing where precision outweighs flow, and for any sentence carrying a quote or a citation.
Business and Casual set the register of the rewrite, so pick the one that matches where the piece will be read. Both are on Starter and up.
Advanced is the heavier rewrite, on Starter and up. The explicit trade-off is that heavier rewrites can flatten authentic voice into a generic conversational register, so use Advanced on individual stubborn sentences only, never as a one-click pass over a whole draft. If a sentence still reads AI after Basic has run twice, Advanced is the last resort; if it still reads AI after Advanced, the underlying claim is probably structurally unfixable and the honest move is to rewrite it by hand.
Scan in 30 seconds. Free, no signup for up to 1,000 words (1,500 with a free account), false-positive rate published; sentence-level highlights on Starter. The verification step of the honest workflow.
Open the detectorThe original-first companion guide for path one writers who want detection to be a side effect of authentic authorship.
Read the guideThe five-step ethical score-reduction workflow for AI-assisted prose. Calibration not shortcuts, with pattern-by-pattern fixes.
Read the guideHow the 0-to-100 score is computed and what thresholds to aim for in general use, graded work, and published prose.
Read the guideFree to try, no card. 3 actions a day (scans or humanizations, any mix), up to 260 words per humanization, one overall AI score on every free scan (sentence-level evidence from Starter), a 90-day audit trail on Pro.
Related walkthroughs on detection, scoring and authentic writing.